DOI: https://doi.org/10.36347/sjet.2026.v14i07.006

Pages: 422-437

Mineral Prospectivity Mapping (MPM) is often constrained by limited labeled samples and the inability of fully supervised graph neural networks to exploit abundant unlabeled spatial data. This study proposes GeoSSGraph, a semi-supervised geologically informed graph network that combines geologically modulated contrastive pretraining (GeoConLoss) with iterative pseudo-labeling (GeoSelfTrain) for mineral prospectivity prediction. The framework integrates Geo Band Weighting and a Geo Relational Encoder comprising three residual basis-decomposed Relational Graph Convolutional Network layers to capture directional geochemical and structural relationships. GeoSSGraph was evaluated on a 43-band geochemical-structural dataset from the Lhasa-Woka region of the eastern Gangdese belt, Tibetan Plateau, consisting of 39 elemental concentration layers, three isometric log-ratio balances, and fracture density. Using a five-model ensemble selected from ten independently trained seeds, GeoSSGraph achieved 89.58% accuracy, 0.9306 ROC-AUC, 0.8980 F1-score, 91.67% sensitivity, and 87.50% specificity on an independent test set. The resulting prospectivity map delineates coherent mineralization corridors aligned with major fault systems and intrusive contacts. These results demonstrate the effectiveness of geologically informed semi-supervised learning for MPM under conditions of label scarcity.